lnenad/lipwise is a speech dictation tool that runs on Windows, macOS, and Linux. Hold down a hotkey and speak, and it goes straight in as a sentence; an AI editor understands spoken instructions like "scratch that" and makes lists or rewrites sentences. Speech recognition is processed on your own computer, and the upper model can be anything from local AI to Claude to OpenAI. The 2026-10-13 commit 814344a pinned the CPU that the built-in speech recognition engine (the ggml inside transcribe-cpp) is built for.
The problem was that the build was machine-relative. ggml defaults to GGML_NATIVE, so the release build simply followed the CPU of the CI runner. The Windows build produced by a runner with AVX-512 died with an illegal instruction on CPUs that do not support that instruction. The commit message cited AMD Ryzen 5800X3D as an example. This patch pins x86 to AVX2·FMA·F16C·BMI2 (Intel Haswell, AMD Zen and newer) and leaves Apple Silicon on the compiler default (M1). The only changed file is src-tauri/.cargo/config.toml, 8 lines added and 0 removed, and it puts GGML_NATIVE=OFF along with the individual options into TRANSCRIBE_CMAKE_ARGS.
That said, this is not a performance improvement but a measure that restores build reproducibility. On CPUs with newer instructions such as AVX-512, the fast path has to be given up, and no figures were given as to how much is lost in that area. A code comment also records that an environment variable of the same name can overwrite this setting. Since the change touches only build configuration, whether already distributed binaries were rebuilt along with it has not been confirmed. What this commit can be said to establish is only that the voice engine no longer depends on the build server's CPU. If you are one of those who experienced the crash, waiting for a new build is the closest thing to a remedy.